Wednesday, September 2nd 2026

Early DLSS 5 Testing Suggests the RTX 5090's Single Power Connector Might Be the Bottleneck

Community testing of leaked DLSS 5 builds suggests the tech may already be hitting a wall on NVIDIA's own flagship GPU. Tom's Hardware ran the neural rendering model through Cyberpunk 2077, Hogwarts Legacy, and Control on an RTX 5090 Founders Edition, as well as an MSI RTX 5090 Lightning Z. Surprisingly, the results point to power delivery as the potential bottleneck rather than compute horsepower. The test used a community Optiscaler injection method built on the leaked DLSS 5 DLL, not NVIDIA's official Streamline path, though the cost lined up closely with NVIDIA's own published 8 ms-per-frame figure.

The Founders Edition, limited to a single 12V-2x6 connector, dropped framerates by 42 to 49 percent with DLSS 5 enabled and consistently hit its 575 W power limit while doing so. The MSI Lightning Z, with a dual-connector setup and a 1000 W limit, took a smaller hit and held better 1% lows, while pulling far more power. In Hogwarts Legacy, it jumped from 480 W to 720 W with neural rendering active, which is a sizeable 50% increase. That leaves single-connector RTX 5090 cards apparently underprovisioned for a feature this demanding, at least unofficially.
The Founders Edition wasn't short on compute so much as it was limited by its own power ceiling. This is notable given how much scrutiny the 12V-2x6 connector has already drawn this generation. DLSS 5 officially launches September 3 with NBA 2K27, and it should become clear within days whether NVIDIA's first-party version fares any better on the Founder's Edition RTX 5090. If not, we might have to wait for future Super variants or next-gen cards to truly maximize the potential of DLSS 5.
Source: Tom's Hardware
Add your own comment

52 Comments on Early DLSS 5 Testing Suggests the RTX 5090's Single Power Connector Might Be the Bottleneck

#26
Tyler-98-W68
TheinsanegamerNNvidia shouldn't ever release a gaming GPU over 75 watt again. All gamers are going to do is whine, cry, and moan about the power connector and power use and woe is me and Nvidia evil ece ece. All those xx6x and higher tier dies would be more profitable as AI processors.

Actually, scratch that, 75watt is the limit of the slot. We can't push connectors to their limit. We need to reduce that. No gaming GPUs over 30 watt.

Nah, let's just ban consumer GPUs. Integrated GPU or bust. Soldered CPUs and memory only. Then there is no melting risk!
Bring back molex connections on gpus
Posted on Reply
#27
Sol_Badguy
So this is the hidden purpose of DLSS5.
To increase power draw.
For the ones without UV to push them into Furmark territory (and mind you Furmark testing lasts far less than gaming sessions).
For the ones with light/moderate UV to get them back at baseline.
No one gets left behind!

Of course the intent had to be disguised into something extremely desirable, this is clearly poison apple maxxing. Jensen has outdone himself.

DLSS5 is the signal that starts the timer, the phone call for the card recall.

Let them come to me, and do not hinder them, for the kingdom of the AI datacenter belongs to such as these.
Posted on Reply
#28
TheinsanegamerN
GrimlingOr you could just design a solid Connector!
See, this is what I'm talking about.

The connector isn't the issue, the lack of load balancing across the pins is. But people will keep repeating the mantra of "the connector, the connector" because that's easier.

So better to just get rid of all external connections, that way there is no problem!
Hecate91600W already pushed the 12v2x6 connector far beyond what it can handle without melting, and it seems DLSS5 will increase the chance of connectors melting even more.
Hardware Unboxed had their 5090 melt, this connector should have never been standardized, or forced upon the entire industry.


Nah that'd be too difficult for a multi-trillion dollar company.
/s of course.
It's not far beyond. The connector is designed to handle up to 625w.

If what you said were true, then every single 5090 would have melted by now. Reality: most don't. Some reddit posts don't change that.
Tyler-98-W68Bring back molex connections on gpus
No those melted too. Failing molex was one of the main drivers behind the PEX 6 pin being developed.

But the 6/8 pins melted too. So really we're just using too much power to play bing bing wahoos. IGPU only is the solution.
Posted on Reply
#29
Hecate91
TheinsanegamerNSee, this is what I'm talking about.

The connector isn't the issue, the lack of load balancing across the pins is. But people will keep repeating the mantra of "the connector, the connector" because that's easier.

So better to just get rid of all external connections, that way there is no problem!
The connector is the issue because load balancing isn't part of the standard.
But as usual people will keep defending this garbage connector. No need to get rid of connectors though, we could simply go back to a connector that worked without gimmicky bandaid fixes such as fancy cables or load sensors, and handled above its rated load.
TheinsanegamerNIt's not far beyond. The connector is designed to handle up to 625w.

If what you said were true, then every single 5090 would have melted by now. Reality: most don't. Some reddit posts don't change that.
Designed and real world are two different things, i'll link you to a nice thread where there are hundreds of reports of the connector melting.
Connectors melting is a well documented issue at the point, no matter if you want to dismiss reddit posts.
Frizz in Close call... 5090 Burnt CableYou wouldn't think it would happen to you until it does... I was reading about 12VHPWR volts on HWINFO and people were saying that you should worry if it's around 11.5 or below at full 600W Load... WH…
TheinsanegamerNNo those melted too. Failing molex was one of the main drivers behind the PEX 6 pin being developed.

But the 6/8 pins melted too. So really we're just using too much power to play bing bing wahoos. IGPU only is the solution.
Those connectors melted less, but of course you'd leave out important context to support your defense.
Posted on Reply
#30
RomeoKnight
YouseethisSo if you push the SPs and rt cores AND tensor cores to the max, you can utilize the "full" chip better resulting in higher power consumption.

I wonder though, if you remove the tensor cores and rt cores and divert that space to rastererization, what would the results be? (at the same power)
Ugly. The result would be ugly as f*ck.
Posted on Reply
#31
3DVCash
mb194dcPushing nearly 1000w for what now ?
As someone who has been a hardware enthusiast since 2010, these numbers honestly feel kinda gross... 😄
Posted on Reply
#32
b1k3rdude
Hecate91600W already pushed the 12v2x6 connector far beyond what it can handle without melting.
Yep, the 8pin molex connector had the PCIe safety factor of 1.92, (288w) or almost double the rated power (150w) for that connector. The 12pin 2x6 has a safety factor of 1.1 or 9%/60w (the connector is only rate for a max of 660w).

While intel/nvidia are ultimately responsible for this shitty connector. PCI-SIG? should have enforced a 400w limit at most to retain the safety factor of 1.92. This would then have retained the the same 100% margin as the old 8pin. But even then the utterly enshitified pin standard compounds the issue. The minute contact points of the female pins and the thinner guage wire, even at the lower 400w power draw its still asking for trouble.
Posted on Reply
#33
Let3r
Sol_BadguySo this is the hidden purpose of DLSS5.
To increase power draw.
For the ones without UV to push them into Furmark territory (and mind you Furmark testing lasts far less than gaming sessions).
For the ones with light/moderate UV to get them back at baseline.
No one gets left behind!

Of course the intent had to be disguised into something extremely desirable, this is clearly poison apple maxxing. Jensen has outdone himself.

DLSS5 is the signal that starts the timer, the phone call for the card recall.

Let them come to me, and do not hinder them, for the kingdom of the AI datacenter belongs to such as these.
Yes! It’s a conspiracy to drive up power prices. Wonder what cut of the profits NVIDIA gets from the power companies. 🤔
Posted on Reply
#34
wolf
Better Than Native
Hecate91
I thought they were
Hecate91Nvidia Unboxed
Shoes on the other foot the moment they sing your song I suppose.
TheinsanegamerNIf what you said were true, then every single 5090 would have melted by now. Reality: most don't. Some reddit posts don't change that.
This is just good for their content cycle like most single topic video's, they'd be silly not to take a jab at anyone they can.
Posted on Reply
#35
ISSEnterprise
Performance drop on DLSS5 even with flagship GPU? Don't worry, just turn down your resolutions a little bit from 4k to 2k etc. Prepare for the incoming RTX6090.

Posted on Reply
#36
AsRock
TPU addict
Yes why stop at one fire.
JoshreddevilIt won't, commoners will purchase time on nvidia's cloud gaming service and be happy.
Yes they will as it be excepted like all the other bs, over time.
Posted on Reply
#37
_roman_
Reading this arcticle made me conclude.

DLSS5 has not 50% penalty, it is even higher.

Higher power consumption and it seems it uses different logic units in the graphic card itself.

You should compare stock card settings vs stock card settings but with power limitations for DLSS5.

I conclude most likely around 70% performance penalty at the same Wattage for DLSS 5.

Undervolt and Overclock etc are advertised features but I consider stock settings as default. And the default settings should be argued with. Just because you bought a more expensive graphic card with selected gpu chip does not imply every chips runs in that setting, etc. I assume, I state without any proof, that some graphic cards modells uses very selected graphic card chips from the board partner.

-- I also conclude any nvidia 5090 graphic card test did not provide solid power consumption figures. Root cause, most likely not all logic units in the core were used.
Posted on Reply
#38
LittleBro
Wow. 42% performance hit for RTX 5090 with DLSS5 slop enabled in Control.
But yes, blame the power limit, not the ineficcient post processing technology that makes NPCs look older and adds brighter tones to ligtning.

Now it's pretty clear why there were 2 RTX 5090s used in Nvidia's demo.
Also, it's pretty clear that running this slop on anything below RTX 5090 will have significantly bigger performance impact.

I'd say DLSS upscaling and FG made sense, since it enabled lower-tier GPUs to render more fps than running on native.
But this technology ... no one asked for it and no one needed it. And since the performance hit is so bad, I wonder who will actually use it.

Anyway, what a discovery was made! When you let GPU draw more power, it will boost to heaven and performance increases! Who would have thought.
Posted on Reply
#39
wolf
Better Than Native
LittleBroBut this technology ... no one asked for it and no one needed it.
I want it and am asking for it. Sorry not sorry. Oops forgot to call it slop :rolleyes:
Posted on Reply
#40
unlox
i'm feeling the more i read about nvidia choices , the more absurd it gets...
Posted on Reply
#41
Youseethis
It looks dramatic but if you ran the numbers the difference is small. Others noticed as well here.

Cyberpunk 2077Cyberpunk 2077 Dlss5DIFF (base/dlss)
FPSPowerFPS/WATTFPSPowerFPS/WATT
5090 fe88.34820.18350.95510.0921.983
5090 Z99.25800.17161.67230.0852.007


ControlControl Dlss5
FPSPowerFPS/WATTFPSPowerFPS/WATT
5090 fe87.55630.15550.95750.0891.756
5090 Z1076910.15561.68020.0772.016




In cyberpunk the difference is tiny between the two cards with the heavily overclocked model being slightly more inefficient overall but utilized slightly better with dlss 5.
In control the 5090 z scales better but it becomes a hog.

Any other 600w model with a large cooler will be adequate to have that little extra oomph for dlss 5 without being a power hog (in comparison)
Posted on Reply
#42
_roman_
more numbers sheet always look like a valid argument, * whataboutism *

does not change the ~10% - too lazy, do not want to think.
0.183 vs 1.171

or the 680 vs 820 on the first post.

break it down in the smallest quantiy, but the ~10% stays. (i do not care, whatever. feel free to be happy it is 10.1%.) ~ 170 + 17 =~~~ i do not care around 183

-- My first impression, I have to revisit an read it again, oh, it is over the ~600 Wattish what should be allowed. hardly any gpu takes the 75Watt from teh slot, so the 675Wattish claims are also invalid

820 Wattish is far more than 600 Wattish what I read over and over again. Note: Whatever if you argue 675Watt, whatever else, 820W is far more than 675W or 600 Watts.

Some run a nvidia 5090 on a daily basis with 1000 Watts, the connector works also with 800 Watts. In some cases for some people, sometimes, for some time.

it seems the hole table is based on Watts and FPS. It should be stated, e.g Power[W] or more common Power - horiontal divider line Watts.

you really did divide fps / power and call it fps/W. bing says: 87.5/563 0.155417407
i would have done it the other way around, Watt / fps -> to see clearly how much watt is needed for a single frame.
luckily bing also has a 1/x feature to show 6.43428570392

691/107 6.45794393

I read the table wrong.

6.43428570392 vs 6.45794393 Watts costs a Frame.

1/x shows 0.154848046195

you have choosen a bad representation for the same data

6.432 vs 6.457 ignoring rounding shows it more clearly (bing calculator website is garbage i do not see a common round function like anyhting had in year 2000)
1/x

When you assume it scales linear, and just take 100 FPS, who cares, 643 vs 645 Watts imply who cares, without rounding.
(note 100 times of these values -> 6.43428570392 vs 6.45794393 Watts costs a Frame.)


darn back button of my mouse .. i had to type it now two times


((575/50.9) / (563/87.5) - 1) * 100 = 75.569762 %

So a DLSS5 frame costs 76% more power around three quarters more power.

I stopped reading igorslab arctice this morning, therefore i should apply the same standards, write a summary statement for the masses. three quarter is better for readability as 76%.

@Youseethis Thanks for sharing your data (you were faster with responding as i could save, etc. my post (mouse button fault not mine, of course it was my fault)

note i usually do this in libreoffice, i can set there the visible digits, i can use round function etc. bing can not to that. I verified the table, one line and i was curious on my statement on it.

-- 76% or 101% (2.016 in your table) is a big penalty per watt.

edit 2. i had the numbesr still in the browser split view. The data seems to be taken from post 1. Post 1 is misleading as hogwards legacy efficiency claim are only 50%, when the other user put in the data and showed clearly 75 to 100% performance penalty per watt. I checked one line the maths, therefore i am certain the other lines are also correct. Using a table and a better approach, clearly show the power consumption doubles in worst case per frame. The efficiency is worse as expected, that surprised me.

I already nuked the buffer, afaik around 6.5Watts vs 11.X Watts per Frame. That says everything. (x = any number from 0 to 9)
Posted on Reply
#43
Youseethis
_roman_more numbers sheet always look like a valid argument, * whataboutism *

does not change the ~10% - too lazy, do not want to think.
0.183 vs 1.171

or the 680 vs 820 on the first post.

break it down in the smallest quantiy, but the ~10% stays. (i do not care, whatever. feel free to be happy it is 10.1%.) ~ 170 + 17 =~~~ i do not care around 183

-- My first impression, I have to revisit an read it again, oh, it is over the ~600 Wattish what should be allowed. hardly any gpu takes the 75Watt from teh slot, so the 675Wattish claims are also invalid

820 Wattish is far more than 600 Wattish what I read over and over again. Note: Whatever if you argue 675Watt, whatever else, 820W is far more than 675W or 600 Watts.

Some run a nvidia 5090 on a daily basis with 1000 Watts, the connector works also with 800 Watts. In some cases for some people, sometimes, for some time.

it seems the hole table is based on Watts and FPS. It should be stated, e.g Power[W] or more common Power - horiontal divider line Watts.

you really did divide fps / power and call it fps/W. bing says: 87.5/563 0.155417407
i would have done it the other way around, Watt / fps -> to see clearly how much watt is needed for a single frame.
luckily bing also has a 1/x feature to show 6.43428570392

691/107 6.45794393

6.43428570392 vs 6.45794393 Watts costs a Frame.

1/x shows 0.154848046195

you have choosen a bad representation for the same data

6.432 vs 6.457 ignoring rounding shows it more clearly (bing calculator website is garbage .. it is not a ti 89)
1/x
The only thing I wanted to show is that both cards behave similarly in cyberpunk. Not what dlss does to efficiency.

I also wanted to see how they scale with power in control where the difference is bigger.

The power connectors will cry and the efficiency falls off a cliff. :roll:
Posted on Reply
#44
mrnagant
Tyler-98-W68Bring back molex connections on gpus
I really liked the Berg connector!
Posted on Reply
#45
Mistral
Is nVidia worried that games don't run slow enough? Is that the play here?
Posted on Reply
#46
LittleBro
MistralIs nVidia worried that games don't run slow enough? Is that the play here?
Nah. The play here is Nvidia must come with other gaming "solutions" rather than releasing Super refresh line or abandoning RTX 6000 series in favor of AI circular money loop.

Unfortunately, AMD will most likely follow, because money, right?
Posted on Reply
#47
Grimling
LittleBroUnfortunately, AMD will most likely follow, because money, right?
I hope AMD is smart enough to see this as an opportunity to crush Nvidia.
If Nvidia cancels the 6000 series and comes up with another 5nm round of the refresh of a refresh, then an 48 GB RDNA5 Card on 3nm will be what people want, because it has electrolytes!
Posted on Reply
#48
Tyler-98-W68
Yes AMD is going to crush Nvidia, they have had so many chances, they have just been bluffing all along...
Posted on Reply
#49
b1k3rdude
ISSEnterprisePerformance drop on DLSS5 even with flagship GPU? Don't worry, just turn down your resolutions a little bit from 4k to 2k etc. Prepare for the incoming RTX6090.
So hardware unboxed has just shown dropping the res/settings doesn't really help. DLSS NR absolutely craters perf from orbit, by anywhere from 2/3 to 3/4 and this is on a 5090.

DLSS NR is scam wrapped in a joke... or a solution looking for a problem... I can only hope game devs have some commons sense and see through this thinly veiled b$ and don't waist any time on it.
Posted on Reply
#50
jnv11
DLSS 5’s extreme power draw made me realize something about VRMs for high end RTX 40 and 50 series GPUs: VRMs that were previously mocked for being overbuilt for those GPUs are actually now looking to be reasonably built for the loads that they will have to handle when DLSS 5 is on, and those that were previously believed to have been reasonably built to match the GPU without being overbuilt now look like they could be undersized.

I am now worried about the GPU and VRM temperatures including their hotspot temperatures when DLSS 5 is enabled.
Posted on Reply
Add your own comment
Nov 22nd, 2026 01:16 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts